<![CDATA[Sentiment analysis is a prominent research area in natural language processing (NLP). For the Indonesian language, IndoBERT has emerged as a leading model due to its competitive performance. However, its effectiveness is strongly influenced by balanced class distribution. A common challenge arises because user reviews on digital platforms, such as the Google Play Store, often exhibit imbalanced classes. This study investigates the effectiveness of the Random Over Sampler (ROS) technique in improving IndoBERT’s performance under imbalanced data conditions. The dataset consists of 13,821 user reviews of the IDN App collected from the Google Play Store between 2015 and July 2025. Prior to modeling, data preprocessing was performed, including punctuation removal, case folding, stopword removal, tokenizing, normalization, and stemming to ensure textual consistency. Reviews were categorized into two sentiment classes: positive (3–5 stars) and negative (1–2 stars). Two experimental scenarios were conducted: (1) IndoBERT without ROS and (2) IndoBERT with a balanced dataset using ROS. Model performance was evaluated using accuracy, precision, recall, and F1-score, with data split into 70% training, 20% validation, and 10% testing. Results showed a significant improvement after ROS implementation: 94.55% accuracy, 94.61% precision, 94.53% recall, and 94.54% F1-score. Confusion matrix analysis indicated improved classification of the minority class, reducing the error rate by 49%. However, learning curve analysis revealed potential overfitting due to ROS. Further research is needed to optimize ROS strategies for better performance and generalization.]]>
